Here is the test hs file https://github.com/rjnw/hakaru-benchmarks/blob/master/testcode/hssrc/nb_simpbucket.hs. I removed pose and categorical, compiled it directly with stack ghc. It takes three command line arguments; location of words, docs, and topics file. I have pushed the news input in the input directory, and also a smaller input news_t which is randomly generated.

I’ll look at this more tomorrow but if you’re not using https://github.com/hakaru-dev/hakaru/commit/a4e93012419b6cfb693d93a7315f7372143cefb8 then please try simplifying with it.

the commit is in S-expression printer! while the problem is in the haskell code. Naive Bayes runs fine in hakrit with full news dataset.